pith. sign in

arxiv: 2605.28226 · v1 · pith:UUV2HCCXnew · submitted 2026-05-27 · 💻 cs.LG

PhAME: Phenotype-Aware Molecular Editing via Latent Diffusion

{\L}ukasz Janisi\'ow , Sebastian Musia{\l} , Bartosz Zieli\'nski , Dawid Rymarczyk , Tomasz Danel This is my paper

Pith reviewed 2026-06-29 13:34 UTC · model grok-4.3

classification 💻 cs.LG
keywords molecular editinglatent diffusionphenotype-aware generationclassifier-free guidancedrug discoverygraph VAEsmall-molecule optimizationcompositional guidance
0
0 comments X

The pith

PhAME edits molecules in a VAE latent space using two independent classifier-free guidance scales to balance phenotypic targets against structural similarity to a seed.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents PhAME as a latent diffusion method that turns molecular optimization into controlled editing inside a pretrained graph VAE. It targets the dual requirement in drug discovery of steering molecules toward specific high-dimensional biological signatures while keeping them close to a known starting structure. The central mechanism is a compositional classifier-free guidance scheme that applies separate scale factors to the phenotype signal and the structural similarity signal. This separation lets users adjust the relative strength of each objective during generation. If the approach works, existing generative models gain a practical way to meet both precision and proximity constraints that prior methods could not handle together.

Core claim

PhAME recasts molecular optimization as editing in the latent space of a pretrained graph-based VAE. Its central contribution is a compositional classifier-free guidance scheme with two independent scales, one for the phenotype-conditioning and one for similarity to the seed structure, allowing practitioners to control the tradeoff between these two objectives. Empirical evaluations across diverse benchmarks, including docking score optimization and multimodal phenotypic generation, demonstrate that PhAME achieves state-of-the-art results while maintaining high chemical validity and novelty.

What carries the argument

Compositional classifier-free guidance scheme with two independent scales, one controlling phenotype-conditioning strength and the other controlling similarity to the seed structure.

If this is right

  • Practitioners gain explicit control over the tradeoff between achieving desired phenotypic signatures and maintaining proximity to a known hit.
  • The method produces state-of-the-art performance on docking score optimization tasks.
  • It supports multimodal phenotypic generation while keeping high chemical validity and novelty.
  • Molecular editing becomes possible without sacrificing either biological relevance or structural closeness to the starting molecule.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar dual-scale guidance could be tested on other latent generative models beyond graph VAEs, such as those operating on SMILES or 3D conformers.
  • The approach may reduce the number of synthesis rounds needed in hit-to-lead campaigns by allowing finer control over how far an edit moves from the original molecule.
  • If the two scales remain independent across different phenotype data types, the framework could support joint optimization of cell morphology and transcriptomic readouts without additional retraining.

Load-bearing premise

The latent space of the pretrained graph-based VAE supports meaningful, controllable edits that simultaneously respect phenotypic signatures and structural proximity.

What would settle it

Generation runs in which increasing the phenotype guidance scale fails to improve phenotypic match while the structure scale fails to preserve seed similarity, or where the two scales cannot be varied independently without one dominating the other.

Figures

Figures reproduced from arXiv: 2605.28226 by Bartosz Zieli\'nski, Dawid Rymarczyk, {\L}ukasz Janisi\'ow, Sebastian Musia{\l}, Tomasz Danel.

Figure 1
Figure 1. Figure 1: Molecular editing with PhAME. Given a seed molecule and target phenotype, the model generates close structural analogs that induce a target phenotype by balancing two independent guidance scales: structure and phenotype. Drug discovery is a search over a chemical space estimated at 1060 drug-like molecules [38], but in practice it rarely starts from scratch. The dom￾inant bottleneck is the iterative refine… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of PhAME. (Training, left) A frozen DGVAE encoder E maps a target molecule to a latent z0, which is noised along a forward diffusion, and a phenotype-aware encoding zalign of a seed molecule along with a target phenotypic signature c is used to train a denoising network. (Inference, right) A molecule is encoded into the latent space and partially noised; then, the denoising network conditioned on … view at source ↗
Figure 3
Figure 3. Figure 3: Influence of guidance scales. Trade-off between target success and similarity under vary￾ing guidance strengths. Increasing property guid￾ance (wc) improves target success, while increas￾ing alignment guidance (wa) enhances similarity to the seed molecule. Notably, PhAME is highly steerable, as illus￾trated in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Top-scoring docked poses of PhAME-generated molecules across five protein targets. For each target, we visualize the highest-affinity molecule generated under docking-score guidance. The protein backbone is shown as cartoon; binding-pocket residues (any heavy-atom contact within 3.5 Å of the ligand) are drawn as light-blue sticks; the docked ligand is in orange. Setup. Following Lee et al. [19], we train o… view at source ↗
Figure 5
Figure 5. Figure 5: Recovery of known binder motifs. Maximum Tanimoto similarity to known binders for each target. PhAME generates compounds most similar to known binders for 8 out of 10 tar￾gets. Bars show mean ± std over 5 runs. Max sim. to binders 0.36 + 39% Seed molecule Generated analog + = HDAC1 0.30 Max sim. to binders + = AKT1 0.57 Max sim. to binders 0.50 Max sim. to binders + 90% [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 6
Figure 6. Figure 6: Gene-driven recovery of known binders. PhAME generates analogs with sub￾stantially increased similarity to known binders using only target-specific knockdown profiles. For AKT1, the maximum Tanimoto similarity to known binders improves by 39%, while for HDAC1 the improvement reaches 90%. Complementary to the structure-based dock￾ing task, we evaluate PhAME on generating compounds conditioned on target-spec… view at source ↗
Figure 7
Figure 7. Figure 7: Phenotype-driven recovery of known drugs. PhAME, run in editing mode and seeded with a structurally similar hit, recovers analogs: aspirin (cyclooxygenase inhibitor, left), clofibrate (PPAR receptor agonist, middle) and dyphylline (adenosine receptor antagonist, right) when condi￾tioned on the corresponding MoA centroid. Results [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of logP in the MOSES Dataset. Distribution of logP values for the training and test sets derived from the MOSES dataset. The dashed vertical line denotes the global median logP of the train set. Model. For this task, we used the c-CFG (Compositional Classifier-Free Guidance) variant of PhAME. Unless otherwise stated, all reported results use wc = 6 and wa = 3. The model was trained for 1000 ep… view at source ↗
Figure 9
Figure 9. Figure 9: Generative Performance Across Training Stages. Novel top 5% docking score and novel hit ratio (%) across training stages. Values denote mean ± standard deviation over 5 runs. Quantitative top-5 hits [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: highlights the advantage of our chosen representation. It shows that CellProfiler achieves the highest overall performance on the Top-1 Cluster Accuracy and MoA Retrieval Rate@k metrics, averaged across all evaluation spaces (CLOOME, InfoAlign, and ECFP) [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Impact of Representations on Controllable Generation. Trade-off between Target Success and Similarity to Seed Molecule under varying guidance strengths for different type encoded representation for conditioning wa. Increasing property guidance (wc) improves target success, while increasing seed guidance (wa) enhances similarity to the input molecule. InfoAlign achieves the highest Target Success, showing … view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative Assessment of PhAME-Generated molecules. Qualitative examples of molecules generated by PhAME, conditioned by MoA embedding centroids. Each row presents a different Mode of Action. The first column captioned as Reference represents one of the selected molecules from our dataset with a known MoA, while the remaining columns visualize the nearest generated molecules based on Tanimoto similarity … view at source ↗
Figure 13
Figure 13. Figure 13: Recovery of Aspirin. Qualitative examples of PhAME recovering Aspirin Lead (Gener￾ated Analog). Six distinct Hits (Seed molecule) were evaluated using structural anchor and biological conditioning of cyclooxygenase inhibitor [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Recovery of Dyphylline. Qualitative examples of PhAME recovering Dyphylline Lead (Generated Analog). Six distinct Hits (Seed molecule) were evaluated using structural anchor and biological conditioning of adenosine receptor antagonist. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Recovery of Clofibrate. Qualitative examples of PhAME recovering Clofibrate Lead (Generated Analog). Six distinct Hits (Seed molecule) were evaluated using structural anchor and biological conditioning of PPAR receptor agonist. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
read the original abstract

Small-molecule drug discovery requires simultaneous optimization of numerous properties of candidate molecules. These properties can be investigated through the analysis of high-dimensional biological signatures, such as cell morphology and transcriptomic perturbations, which provide a rich perspective on the underlying biological mechanisms. However, existing generative methods, which use those signatures for optimization, fail to meet two key requirements: providing precise guidance toward desired phenotypic signatures while maintaining structural proximity to a known hit. We introduce PhAME (Phenotype-Aware Molecular Editing), a latent diffusion framework that overcomes this challenge by recasting molecular optimization as editing in the latent space of a pretrained graph-based VAE. Our central contribution is a compositional classifier-free guidance scheme with two independent scales, one for the phenotype-conditioning and one for similarity to the seed structure, allowing practitioners to control the tradeoff between these two objectives. Empirical evaluations across diverse benchmarks, including docking score optimization and multimodal phenotypic generation, demonstrate that PhAME achieves state-of-the-art results while maintaining high chemical validity and novelty.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces PhAME, a latent diffusion framework for molecular editing that recasts optimization as editing in the latent space of a pretrained graph-based VAE. Its central contribution is a compositional classifier-free guidance scheme using two independent scales—one for phenotype conditioning and one for similarity to the seed structure—to control the tradeoff between phenotypic signatures and structural proximity. The work claims state-of-the-art empirical results on docking score optimization and multimodal phenotypic generation benchmarks while preserving high chemical validity and novelty.

Significance. If the two-scale guidance scheme successfully enables independent control without feature entanglement, the method would offer a practical advance for multi-objective small-molecule optimization in drug discovery by allowing tunable tradeoffs between biological phenotype matching and retention of known hit structures. The approach leverages standard classifier-free guidance in a compositional manner on top of an existing VAE, which is a reasonable extension, but its impact hinges on whether the frozen VAE latent space actually supports the required separation.

major comments (1)
  1. [Abstract] Abstract: The central claim that the compositional classifier-free guidance with two independent scales allows controllable tradeoffs presupposes that the pretrained graph VAE latent space contains factors that can be modulated separately for phenotypic signatures versus structural proximity to the seed. Standard graph VAEs are trained solely on molecular graphs without phenotype supervision, so phenotype information is likely entangled with structural features; the manuscript provides no explicit mechanism, loss term, or post-hoc analysis to guarantee orthogonality or independent controllability of the two conditioning signals once the diffusion model is trained on the frozen VAE. This assumption is load-bearing for the claimed advantage over prior generative methods.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying the key assumption underlying our central contribution. We respond to the major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the compositional classifier-free guidance with two independent scales allows controllable tradeoffs presupposes that the pretrained graph VAE latent space contains factors that can be modulated separately for phenotypic signatures versus structural proximity to the seed. Standard graph VAEs are trained solely on molecular graphs without phenotype supervision, so phenotype information is likely entangled with structural features; the manuscript provides no explicit mechanism, loss term, or post-hoc analysis to guarantee orthogonality or independent controllability of the two conditioning signals once the diffusion model is trained on the frozen VAE. This assumption is load-bearing for the claimed advantage over prior generative methods.

    Authors: We agree that the pretrained graph VAE is unsupervised with respect to phenotype and that the manuscript does not supply an explicit loss term or theoretical guarantee of orthogonality between phenotypic and structural factors in the latent space. The method instead demonstrates practical controllability through the compositional classifier-free guidance applied at sampling time: the diffusion model is trained to predict noise under both phenotype and seed-structure conditions, after which independent guidance scales are used to steer the two objectives. Our empirical results across docking-score optimization and multimodal phenotypic benchmarks show that varying the two scales produces the expected tradeoffs in phenotypic fidelity versus structural similarity while preserving validity. We acknowledge the absence of post-hoc analysis of latent-space entanglement and will add, in revision, both a discussion of this assumption and new experiments that quantify correlations between phenotype predictions and structural features in the VAE latent space together with further guidance-scale ablations. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained methodological contribution

full rationale

The paper introduces a latent diffusion model with a new compositional classifier-free guidance scheme using two independent scales. No equations, fitting procedures, or self-citations are visible that reduce the central claim to an input by construction. The method is framed as a novel recasting of optimization as editing in a pretrained VAE latent space, with empirical evaluations presented as external validation. This matches the default expectation of no circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no equations, training details, or explicit assumptions; ledger is therefore empty.

pith-pipeline@v0.9.1-grok · 5721 in / 974 out tokens · 27332 ms · 2026-06-29T13:34:43.339921+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 6 canonical work pages

  1. [1]

    Alhossary, S

    A. Alhossary, S. D. Handoko, Y . Mu, and C.-K. Kwoh. Fast, accurate, and reliable molecular docking with quickvina 2.Bioinformatics, 31(13):2214–2216, 2015

    2015

  2. [2]

    M.-A. Bray, S. Singh, H. Han, C. T. Davis, B. Borgeson, C. Hartland, M. Kost-Alimova, S. M. Gustafsdottir, C. C. Gibson, and A. E. Carpenter. Cell painting, a high-content image- based assay for morphological profiling using multiplexed fluorescent dyes.Nature protocols, 11(9):1757–1774, 2016

    2016

  3. [3]

    A. E. Carpenter, T. R. Jones, M. R. Lamprecht, C. Clarke, I. H. Kang, O. Friman, D. A. Guertin, J. H. Chang, R. A. Lindquist, J. Moffat, P. Golland, and D. M. Sabatini. Cellprofiler: image analysis software for identifying and quantifying cell phenotypes. 2006

    2006

  4. [4]

    S. N. Chandrasekaran, J. Ackerman, E. Alix, D. M. Ando, J. Arevalo, M. Bennion, N. Boisseau, A. Borowa, J. D. Boyd, and e. a. Brino. Jump cell painting dataset: morphological impact of 136,000 chemical and genetic perturbations.bioRxiv, 2023

    2023

  5. [5]

    S. N. Chandrasekaran, H. Ceulemans, J. D. Boyd, and A. E. Carpenter. Image-based profiling for drug discovery: due for a machine-learning upgrade?Nature reviews drug discovery, 20(2):145–159, 2021

    2021

  6. [6]

    S. M. Corsello, J. A. Bittker, Z. Liu, J. Gould, P. McCarren, J. E. Hirschman, S. E. Johnston, A. Vrcic, B. Wong, M. Khan, J. Asiedu, R. Narayan, C. C. Mader, A. Subramanian, and T. R. Golub. The drug repurposing hub: a next-generation drug library and information resource. Nature medicine, 2017. 10

    2017

  7. [7]

    Q. Duan, C. Flynn, M. Niepel, M. Hafner, J. L. Muhlich, N. F. Fernandez, A. D. Rouillard, C. M. Tan, E. Y . Chen, T. R. Golub, et al. Lincs canvas browser: interactive web app to query, browse and interrogate lincs l1000 gene expression signatures.Nucleic acids research, 42(W1):W449–W460, 2014

    2014

  8. [8]

    Gaulton, L

    A. Gaulton, L. J. Bellis, A. P. Bento, J. Chambers, M. Davies, A. Hersey, Y . Light, S. McGlinchey, D. Michalovich, B. Al-Lazikani, and J. P. Overington. Chembl: a large-scale bioactivity database for drug discovery.Nucleic Acids Research, 2012

    2012

  9. [9]

    Gómez-Bombarelli, J

    R. Gómez-Bombarelli, J. N. Wei, D. Duvenaud, J. M. Hernández-Lobato, B. Sánchez-Lengeling, D. Sheberla, J. Aguilera-Iparraguirre, T. D. Hirzel, R. P. Adams, and A. Aspuru-Guzik. Auto- matic chemical design using a data-driven continuous representation of molecules.ACS central science, 4(2):268–276, 2018

    2018

  10. [10]

    Gupta, B

    A. Gupta, B. Lenin, S. Current, R. Batra, B. Ravindran, K. Raman, and S. Parthasarathy. Pure: policy-guided unbiased representations for structure-constrained molecular generation.Journal of Cheminformatics, 17(1):1–18, 2025

    2025

  11. [11]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, editors,Advances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Associates, Inc., 2020

    2020

  12. [12]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance, 2022

    2022

  13. [13]

    Hoogeboom, V

    E. Hoogeboom, V . G. Satorras, C. Vignac, and M. Welling. Equivariant diffusion for molecule generation in 3d. InInternational conference on machine learning, pages 8867–8887. PMLR, 2022

    2022

  14. [14]

    J. J. Irwin and B. K. Shoichet. Zinc- a free database of commercially available compounds for virtual screening.Journal of chemical information and modeling, 45(1):177–182, 2005

    2005

  15. [15]

    Jiang, X

    Q. Jiang, X. Ye, Z. Guo, Y . Xia, Z. Liu, J. Xu, P. Jin, F. Ju, H. Xia, S. Feng, et al. Phenotype- guided in silico molecular generation using large language models.bioRxiv, pages 2026–01, 2026

    2026

  16. [16]

    W. Jin, R. Barzilay, and T. Jaakkola. Junction tree variational autoencoder for molecular graph generation. InInternational conference on machine learning, pages 2323–2332. PMLR, 2018

    2018

  17. [17]

    B. Jing, G. Corso, J. Chang, R. Barzilay, and T. Jaakkola. Torsional diffusion for molecular conformer generation.Advances in neural information processing systems, 35:24240–24253, 2022

    2022

  18. [18]

    Kaitoh and Y

    K. Kaitoh and Y . Yamanishi. Triomphe: Transcriptome-based inference and generation of molecules with desired phenotypes by machine learning.Journal of Chemical Information and Modeling, 61(9):4303–4320, 2021

    2021

  19. [19]

    S. Lee, J. Jo, and S. J. Hwang. Exploring chemical space with score-based out-of-distribution generation. InInternational Conference on Machine Learning, pages 18872–18892. PMLR, 2023

    2023

  20. [20]

    S. Lee, K. Kreis, S. P. Veccham, M. Liu, D. Reidenbach, Y . Peng, S. Paliwal, W. Nie, and A. Vahdat. Genmol: A drug discovery generalist with discrete diffusion.arXiv preprint arXiv:2501.06158, 2025

    work page arXiv 2025

  21. [21]

    Li and Y

    C. Li and Y . Yamanishi. Gxvaes: Two joint vaes generate hit molecules from gene expression profiles. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 13455–13463, 2024

    2024

  22. [22]

    Li and Y

    C. Li and Y . Yamanishi. Gxvaes: Two joint vaes generate hit molecules from gene expression profiles.Proceedings of the AAAI Conference on Artificial Intelligence, 38(12):13455–13463, Mar. 2024

    2024

  23. [23]

    scppdm: A diffusion model for single-cell drug-response prediction.arXiv preprint arXiv:2510.11726, 2025

    Z. Liang, S. Zhuang, X. Jiao, W. Mao, H. Chen, and C. Shen. scppdm: A diffusion model for single-cell drug-response prediction.arXiv preprint arXiv:2510.11726, 2025. 11

    work page arXiv 2025

  24. [24]

    G. Liu, S. Seal, J. Arevalo, Z. Liang, A. E. Carpenter, M. Jiang, and S. Singh. Learning molecular representation in a cell. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  25. [25]

    H. Liu, S. Tian, and X. Liu. Phenotypic profile-informed generation of drug-like molecules via dual-channel variational autoencoders.arXiv preprint arXiv:2506.02051, 2025

    work page arXiv 2025

  26. [26]

    H. Liu, S. Tian, and X. Liu. Phenotypic profile-informed generation of drug-like molecules via dual-channel variational autoencoders. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI ’25, 2025

    2025

  27. [27]

    A. A. Lobo, U. Awasthi, and L. Zhukov. Property-guided molecular generation and optimization via latent flows.arXiv preprint arXiv:2603.26889, 2026

    work page arXiv 2026

  28. [28]

    H. H. Loeffler, J. He, A. Tibo, J. P. Janet, A. V oronov, L. H. Mervin, and O. Engkvist. Reinvent 4: Modern ai–driven generative molecule design.Journal of Cheminformatics, 16(1):20, 2024

    2024

  29. [29]

    S. Z. Lu, Z. Lu, E. Hajiramezanali, T. Biancalani, Y . Bengio, G. Scalia, and M. Koziarski. Cell morphology-guided small molecule generation with gflownets. InICML 2024 Workshop on Structured Probabilistic Inference{\&}Generative Modeling, 2024

    2024

  30. [30]

    Maziarka, A

    Ł. Maziarka, A. Pocha, J. Kaczmarczyk, K. Rataj, T. Danel, and M. Warchoł. Mol-cyclegan: a generative model for molecular optimization.Journal of Cheminformatics, 12(1):2, 2020

    2020

  31. [31]

    Méndez-Lucio, B

    O. Méndez-Lucio, B. Baillif, D.-A. Clevert, D. Rouquié, and J. Wichard. De novo generation of hit-like molecules from gene expression signatures using artificial intelligence.Nature communications, 11(1):10, 2020

    2020

  32. [32]

    F. M. Morelli, V . Kim, F. Hecker, S. Geibel, and P. A. Marín Zapata. unidino: Assay-independent feature extraction for fluorescence microscopy images.Computational and Structural Biotech- nology Journal, 27:928–936, 2025

    2025

  33. [33]

    Moshkov, M

    N. Moshkov, M. Bornholdt, S. Benoit, M. Smith, C. McQuin, A. Goodman, R. A. Senft, Y . Han, M. Babadi, P. Horvath, B. A. Cimini, A. E. Carpenter, S. Singh, and J. C. Caicedo. Learning representations for image-based profiling of perturbations.bioRxiv, 2022

    2022

  34. [34]

    Nguyen and A

    T. Nguyen and A. Karolak. Expanding molecular design with graph variational autoencoders: A comparative study of pair-encoding and character tokenization.ChemRxiv, 2025(0523), 2025

    2025

  35. [35]

    S. K. Niazi. The coming of age of ai/ml in drug discovery, development, clinical testing, and manufacturing: the fda perspectives.Drug Design, Development and Therapy, pages 2691–2725, 2023

    2023

  36. [36]

    A. Q. Nichol and P. Dhariwal. Improved denoising diffusion probabilistic models. InInterna- tional conference on machine learning, pages 8162–8171. PMLR, 2021

    2021

  37. [37]

    International Conference on Learning Representations , year=

    A. Palma, T. Richter, H. Zhang, M. Lubetzki, A. Tong, A. Dittadi, and F. Theis. Multi-modal and multi-attribute generation of single cells with cfgen.arXiv preprint arXiv:2407.11734, 2024

    work page arXiv 2024

  38. [38]

    P. G. Polishchuk, T. I. Madzhidov, and A. Varnek. Estimation of the size of drug-like chemical space based on gdb-17 data.Journal of computer-aided molecular design, 27(8):675–679, 2013

    2013

  39. [39]

    Polykovskiy, A

    D. Polykovskiy, A. Zhebrak, B. Sanchez-Lengeling, S. Golovanov, O. Tatanov, S. Belyaev, R. Kurbanov, A. Artamonov, V . Aladinskiy, M. Veselov, et al. Molecular sets (moses): a bench- marking platform for molecular generation models.Frontiers in pharmacology, 11:565644, 2020

    2020

  40. [40]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  41. [41]

    Sanchez-Fernandez, E

    A. Sanchez-Fernandez, E. Rumetshofer, S. Hochreiter, and G. Klambauer. Cloome: contrastive learning unlocks bioimaging databases for queries with chemical structures. 2023. 12

    2023

  42. [42]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021

    2021

  43. [43]

    Song and H

    R. Song and H. Liu. Phenomoler: Phenotype-guided molecular optimization via chemistry large language model.arXiv preprint arXiv:2509.21424, 2025

    work page arXiv 2025

  44. [44]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations, 2021

    2021

  45. [45]

    Subramanian, R

    A. Subramanian, R. Narayan, S. M. Corsello, D. D. Peck, T. E. Natoli, X. Lu, J. Gould, J. F. Davis, A. A. Tubelli, J. K. Asiedu, D. L. Lahr, J. E. Hirschman, Z. Liu, M. Donahue, B. Julian, M. Khan, D. Wadden, I. C. Smith, D. Lam, A. Liberzon, C. Toder, M. Bagul, M. Orzechowski, O. M. Enache, F. Piccioni, S. A. Johnson, N. J. Lyons, A. H. Berger, A. F. Sha...

    2017

  46. [46]

    D. C. Swinney. Phenotypic vs. target-based drug discovery for first-in-class medicines.Clinical Pharmacology & Therapeutics, 93(4):299–301, 2013

    2013

  47. [47]

    Q. Tang, D. Ding, X. Yuan, G. Seabra, P. A. Ramdhan, C.-Y . Liu, M. T. Thai, C. Li, H. Luesch, and Y . Li. Mgmg: Cell morphology-guided molecule generation for drug discovery.bioRxiv, 2025

    2025

  48. [48]

    Vignac, I

    C. Vignac, I. Krawczuk, A. Siraudin, B. Wang, V . Cevher, and P. Frossard. Digress: Discrete denoising diffusion for graph generation. InProceedings of the 11th International Conference on Learning Representations, 2023

    2023

  49. [49]

    C. Wang, H. H. Ong, S. Chiba, and J. C. Rajapakse. Gldm: hit molecule generation with constrained graph latent diffusion model.Briefings in bioinformatics, 25(3):bbae142, 2024

    2024

  50. [50]

    Weiss, E

    T. Weiss, E. Mayo Yanes, S. Chakraborty, L. Cosmo, A. M. Bronstein, and R. Gershoni-Poranne. Guided diffusion for inverse molecular design.Nature Computational Science, 3(10):873–882, 2023

    2023

  51. [51]

    S. Yang, D. Hwang, S. Lee, S. Ryu, and S. J. Hwang. Hit and lead discovery with explorative rl and fragment-based molecule generation.Advances in Neural Information Processing Systems, 34:7924–7936, 2021

    2021

  52. [52]

    P. A. M. Zapata, O. Méndez-Lucio, T. Le, C. J. Beese, J. Wichard, D. Rouquié, and D.-A. Clevert. Cell morphology-guided de novo hit design by conditioning gans on phenotypic image features.Digital discovery, 2(1):91–102, 2023

    2023

  53. [53]

    C. Zeng, J. Jin, C. Ambrose, G. Karypis, M. Transtrum, E. B. Tadmor, R. G. Hennig, A. Roitberg, S. Martiniani, and M. Liu. Propmolflow: property-guided molecule generation with geometry- complete flow matching.Nature Computational Science, pages 1–10, 2026

    2026

  54. [54]

    Zhang, J

    Q. Zhang, J. Xiao, D. Niu, Z. Zhang, S. Ding, and Z. Li. Geometry-complete latent diffusion model for 3d molecule generation.Bioinformatics, 41(8):btaf426, 2025. 13 A Training Details Python environment.All experiments were carried out using Python 3.11.5. The key libraries are listed in Table 5. Other requirements are included in the code repository. Tab...

    2025